Abstract
Animals construct internal "cognitive maps" of the world during navigation in spatial and non-spatial domains, with grid cells in the medial entorhinal cortex (MEC) playing a key role. This requires associating internal position estimates with external cues to reduce spatial uncertainty over time. However, how grid cell representations evolve in novel spaces to support map formation is unclear. To address this question, we longitudinally imaged calcium dynamics of grid cells over 10 days as mice learn operant tasks in novel virtual linear tracks. We observe that spatial tuning of grid cells is present immediately in novel tracks but evolves as a significant fraction of spatial fields shift backward on a run-by-run basis, within and across days. Backward shifts are more prevalent and persistent in successful learners. The fields gradually stabilize across days, anchored by landmarks, suggesting slow plasticity. The backward shifts partially reset daily, reflecting a slower consolidation timescale. While individual fields of a cell shift differentially, co-active fields of co-modular grid cells shift together, indicating their coupled dynamics on the same two-dimensional torus. Spatial learning leads to systematic changes and stabilization of their population phase trajectory, including lateral shift, rotation, and phase resets at landmarks, forming a landmark-fragmented representation for the environment. Next, we build an entorhinal-hippocampal model that provides a mechanistic explanation of the diverse phenomena - grid field shifts, increasing fidelity, and fragmentation of the spatial map - and predicts slow Hebbian plasticity in the hippocampus-to-entorhinal pathway. Supporting this, electrophysiology demonstrates that learning-performance-correlated weakening of local inhibition facilitates potentiation of indirect hippocampal inputs to superficial MEC. Together, our study provides multifaceted evidence of slow hippocampus-to-MEC plasticity, elucidating the formation of stable and fragmented cognitive maps that combine internal and cue-driven positional estimates in rich environments during learning. This mechanism may extend to broader memory processes involving this circuit.